tracel-ai/burn · error

no cxx11 abi returned by python {output:?}

Error message

no cxx11 abi returned by python {output:?}

What it means

When LIBTORCH_USE_PYTORCH is set, burn-tch's build script queries the active Python interpreter to locate the installed PyTorch/libtorch and to determine the C++11 ABI flag. If the python probe output contains no `LIBTORCH_LIB: ` line (so no library dir) and/or no cxx11 ABI answer, the script panics. Essentially: python found PyTorch but didn't report the metadata the build expects.

Source

Thrown at crates/burn-tch/build.rs:91

                .output()
                .expect("error running python interpreter");
            let mut cxx11_abi = None;
            for line in String::from_utf8_lossy(&output.stdout).lines() {
                match line.strip_prefix("LIBTORCH_CXX11: ") {
                    Some("True") => cxx11_abi = Some("1".to_owned()),
                    Some("False") => cxx11_abi = Some("0".to_owned()),
                    _ => {}
                }
                if let Some(path) = line.strip_prefix("LIBTORCH_INCLUDE: ") {
                    libtorch_include_dirs.push(PathBuf::from(path))
                }
                if let Some(path) = line.strip_prefix("LIBTORCH_LIB: ") {
                    libtorch_lib_dir = Some(PathBuf::from(path))
                }
            }
            match cxx11_abi {
                Some(cxx11_abi) => cxx11_abi,
                None => panic!("no cxx11 abi returned by python {output:?}"),
            }
        } else {
            let libtorch = Self::prepare_libtorch_dir(os)?;
            let includes = env_var_rerun("LIBTORCH_INCLUDE")
                .map(PathBuf::from)
                .unwrap_or_else(|_| libtorch.clone());
            let lib = env_var_rerun("LIBTORCH_LIB")
                .map(PathBuf::from)
                .unwrap_or_else(|_| libtorch.clone());
            libtorch_include_dirs.push(includes.join("include"));
            libtorch_include_dirs.push(includes.join("include/torch/csrc/api/include"));
            if lib.ends_with("lib") {
                // DEP_TCH_LIBTORCH_LIB might already point to /lib
                libtorch_lib_dir = Some(lib);
            } else {
                libtorch_lib_dir = Some(lib.join("lib"));
            }
            env_var_rerun("LIBTORCH_CXX11_ABI").unwrap_or_else(|_| "1".to_owned())

View on GitHub (pinned to d16f7ba2ed)

Solutions

1. Activate/point to the Python env that has torch installed before building (e.g. `source venv/bin/activate` or set PYTHON_SYS_EXECUTABLE).
2. Verify `python -c "import torch; print(torch.__version__)"` works in the build shell; install torch if missing.
3. Pin a PyTorch version compatible with your tch/burn-tch versions (check tch's supported libtorch versions) and rebuild.
4. Alternatively skip the python probe: set LIBTORCH / LIBTORCH_LIB and LIBTORCH_INCLUDE env vars to a downloaded libtorch distribution.

Example fix

// before (build shell)
LIBTORCH_USE_PYTORCH=1 cargo build --features tch   # 'python' has no torch -> panic
// after
source ~/miniconda3/envs/ml/bin/activate            # env where torch is installed
LIBTORCH_USE_PYTORCH=1 cargo build --features tch

# run before cargo build with LIBTORCH_USE_PYTORCH=1
import torch, sys
try:
    import torch
    print(torch.__version__)
except ImportError:
    sys.exit("active python has no torch; activate the right env or set LIBTORCH manually")

When it happens

Trigger: Building with LIBTORCH_USE_PYTORCH=1 when `python` prints unexpected output — wrong interpreter on PATH without torch, a conda/venv mismatch, a torch version whose probe script output format differs, or pip-installed CPU/GPU torch layouts the probe can't parse.

Common situations: Multiple Python installs (system vs conda vs pyenv) so `python` isn't the one with torch installed; recently upgraded PyTorch to a version incompatible with the pinned tch crate; missing python3-torch distro package; CI image where torch is installed for a different interpreter.

Related errors

• error running python interpreter
• LibTorch is not supported as a remote-server backend (no Bac
• unsupported TARGET_OS '{os}'
• Unsupported dtype for `bool_from_data`
• Unsupported dtype for `int_from_data`: {:?}

AI-assisted analysis of tracel-ai/burn@d16f7ba2ed (2026-09-05). Data as JSON: /api/errors/44e9ce20054bd099. Report an issue: GitHub.